Vibe coding and AI agents classes in Huyton with Roby

The learner sets the agent a placement puzzle. The Census output areas that make up Huyton with Roby number 193 within Knowsley, holding 58,998 residents by our count, and the ONS publishes a population-weighted centre point for each. Six sites must be put on six of those points so that the average resident is as close as possible to one of them, measured in a straight line. There are far too many layouts to try them all. So the agent starts from six random points and improves: a move shifts one site to one of the eight areas closest to it.

Plain local search takes whichever move helps most and stops when none helps. Tabu search never stops early. At every turn it takes the least harmful move available, even when that move makes the layout worse, and it writes the area it has just left onto a tabu list so that it cannot step straight back. The length of that list is the agent's memory. Both methods were run from the same 40 random starts.

Local search ends within 1% of 573.0 metres, the lowest figure any run reached, in 31 of the 40 attempts; the other nine are dead ends, the worst at 597.2. Letting the agent move downhill without a memory changes almost nothing, because it steps out of the dead end and straight back in. A memory of ten moves brings all 40 runs within 1%. A memory of 60 does slightly worse again, since so many moves are forbidden that good ones are blocked. The price is effort: about 7,199 layouts scored per run against a median of 1,072 for local search. And 573.0 is only the lowest we found. Nothing here proves a better layout does not exist.

Teenagers in Huyton who can code a memory into an agent understand why an assistant goes round in circles, and that understanding comes from writing the code themselves. The longer argument is in why teenagers should still learn to code in 2026.